Your customer asks about your AI. Here is how to answer.

More and more buyers add AI questions to their security and vendor reviews. Saying your AI is safe is easy. Showing that it holds is what matters.

Questions to expect

• Which AI models do you use, and from which provider?
• Is our data used to train models?
• Can the AI show one customer’s data to another customer?
• How do you stop the AI from making things up or ignoring instructions?
• When does the AI hand over to a human?
• Do users know they are talking to an AI?

Why that last question comes up more often

Under Article 50 of the EU AI Act, AI systems that interact with people must make clear that they are AI. That obligation has applied since 2 August 2026. Customers want to know whether your product complies.

Documents are not enough

For many questions, policies or architecture are enough. But a claim such as “our AI never shares another customer’s data” is about behaviour. Whether it holds only shows when you test it, with realistic and adversarial requests, and more than once, because an AI does not answer the same way every time.

How to approach it

1. Make your claims precise. Turn “our AI is safe” into statements you can check.
2. Use documents for what documents can prove. Policies, contracts with model providers, architecture.
3. Test what is behaviour. In a test environment, with fictional data, and record what you tested and when.
4. Be honest about the outcome. If you find a problem, fix it and retest before your customer finds it.

How Holdvero helps

With AI Deal Assurance we complete your security and AI questionnaire with traceable evidence, translate your key AI claims into testable assertions and test them in your staging environment. You get a report for your customer with the outcome per claim, the test configuration and run IDs, plus one retest after remediation. Price: €3,500, delivered 5–10 working days after complete intake and test access. Not a certification, real evidence.